Gokul Rajaram on the 8 Moats Companies Need & Why Dropouts are "AI Maxing" the World
Summary
Gokul Rajaram sees the indiscriminate software selloff as a “100% overreaction,” but cheap code has raised the bar for defensibility. His eight-moat test scores proprietary data, embedded workflow, regulation, exclusive distribution, ecosystem, network effects, physical infrastructure, and scale; four or more makes a company “pretty damn secure,” two or three is weak, and zero means “you’re screwed.”
Enduring companies pair a remarkable core product with distribution and a naturally adjacent multi-product portfolio. Google’s internal Caribou project offered 1 GB of email storage against Yahoo Mail’s 10 MB; Facebook demonstrated why multiplayer products distribute and defend themselves; and Square grew from payments into 11 products exceeding $50 million of revenue each, by Rajaram’s recollection. Crucially, some products own the profit pool while others exist to improve retention.
For early-stage pure software, defensibility largely collapses to two questions: does the proprietary data compound, and how deeply does the product control the workflow? Rajaram put Atlassian at about three moats and described Salesforce as similar, while saying Monday might be rightly priced in this environment. He argues systems of record must “commoditize the complement”: charge for either valuable workflows or data, while giving the other away before agent companies capture the profit pool.
A narrow vertical agent may be viable, but a venture-scale vertical company must own the full stack and ultimately attack labor spend. ServiceTitan had roughly 32 products yet, as Rajaram put it, was still a sub-$10 billion company or something like that, while horizontal platforms such as Robinhood and Coinbase have 13 and 12 $100 million-plus product lines, respectively. AI enters enterprise budgets first by replacing outsourced BPO at 20–30% lower cost, then by preventing backfills, and only later through layoffs.
Explosive growth is no longer enough: durability is the scarce signal. “One to 10” has become common, while Jasper’s rapid rise and reversal illustrates the danger of tire-kickers; Rajaram would prefer triple-triple-double-double growth with excellent gross and net retention to 10x growth with sub-90% net revenue retention. Even strong cohorts must be tested against a “seismic event” such as a credible bundled competitor.
Margins should be underwritten through future pricing power, while valuation discipline depends heavily on stage. Falling inference costs should improve gross margins, but Rajaram prefers businesses capable of raising prices because they have durable leverage over customers; at seed or Series A, an exceptional outcome can overwhelm entry price, whereas at Series B and beyond a good company bought at $4 billion can grow revenue from $100 million to $500 million and still produce no return.
Rajaram has reversed his belief that fully remote early-stage companies can scale, after watching founding teams fail to align and companies die despite otherwise promising ingredients. He now wants at least three in-person days a week. He simultaneously urges most graduates to gain two or three years of operating experience, while acknowledging that exceptional young founders are unusually “AI maxed” and that he has backed more dropouts recently than in his previous 15 years combined.
Deep dive
1. Remarkability starts the company; distribution and product breadth make it endure
Google taught Rajaram that no amount of go-to-market rescues an unremarkable product. His test is whether the core value proposition is genuinely “10x, 100x better than the alternative,” not merely packaged or sold more effectively.
His defining example was Caribou, Google’s internal webmail project: 1 GB of free storage when Yahoo Mail offered 10 MB. Released, as he recalled, on April 1, 2003, it looked so implausible that users assumed it was an April Fools’ joke.
Facebook supplied the missing lesson: even remarkable products need distribution. Zuckerberg’s particular genius was recognizing why products would fail to spread, while multiplayer products such as Facebook and later Figma created sharing, switching, and defensibility that single-player software lacked.
Square showed why one product is insufficient. It went from payments alone to, Rajaram thinks, 11 products above $50 million in revenue each; median products used per merchant became a North Star because adoption increased retention. Square Capital barely contributed profit but deepened loyalty, illustrating why teams must distinguish “profit pool products” from “retentive products.”
2. The eight-moat scorecard separates damaged software from doomed software
Rajaram’s first pair is proprietary data and workflow. Spotify’s decade of listening behavior supports a Discover experience that cannot simply be recreated; workflow strength depends on depth, making NetSuite’s business-running ERP closer to a full point while Zendesk might merit only half.
Regulation and distribution create barriers code generation cannot erase. Coinbase has state money-transmission licenses and FinCEN registration; Intuit’s grip on accountants forced Rajaram to abandon Xero for QuickBooks because his accountant simply said, “I’m sorry. I don’t use Xero.”
Ecosystem and network moats live outside the codebase. AI might reproduce Shopify’s storefront software but not the hundreds of thousands of developers and third parties around it; likewise, it cannot reproduce DoorDash’s restaurant access, courier density, liquidity, and reputation history.
Physical infrastructure and scale complete the eight. “Wherever you have atoms,” displacement is harder, while Amazon and TSMC illustrate costs that competitors cannot readily match. Rajaram’s scoring rule: four-plus moats is “pretty damn secure”; two or three is weak; one demands more moat-building; zero means “you’re screwed.”
3. Atlassian may be oversold; Salesforce must choose its profit pool
Applying the framework, Rajaram said Atlassian was being massively oversold, while Monday might be rightly priced in this environment. He later described Salesforce as similar to Atlassian, which he put at a score of about three.
Rajaram did not think Shopify would build Klaviyo, saying—in his opinion—that Shopify had decided the product was not part of its mission. Klaviyo’s defense therefore depends on whether Shopify provides genuinely privileged distribution: if it is the preferred communications product in Shopify’s ecosystem, that part of the business may be difficult to displace.
Harry challenged Rajaram’s exclusion of brand and argued that switching costs could approach zero as data portability and pixel-level experience replication improve. Rajaram disagreed for business software: enterprise buyers are more rational, clones will become stronger, and brand will therefore be weaker on the business side.
Salesforce has about three moats, Rajaram said—workflow, distribution, and ecosystem, but not scale, because cheap software production removes the old scale advantage. Its strategic choice is stark: if workflows hold the profit pool, make data storage free; if data holds it, give agentic workflows away. “They have to commoditize the complement.”
4. Successful AI additions rebuild the experience around new model capabilities
A bolt-on AI feature has “a real ceiling.” The useful distinction is between adding AI search to an old interface and rebuilding search around new UX primitives; the former is an upgrade, while the latter changes what the product actually does.
Rajaram thought Notion’s agents were promising, but said they must learn from customer interactions and be tuned for Notion’s user base. Merely wrapping a GPT or Anthropic model leaves a thin layer; the company must rebuild the experience end to end and accumulate an improving data asset.
Document processing shows the required reset. Capabilities that could not reliably extract structure from unstructured documents six or nine months earlier can now parse dense legal contracts, so an upload flow should immediately infer and surface what is happening. With material model advances every six months, long product roadmaps risk being “blown out by the next model iteration.”
5. Pure software hangs on compounding data, embedded workflow, and shipping velocity
Physical, network, ecosystem, and scale advantages are usually unavailable or unknowable at seed. Distribution hacks rarely endure, leaving investors to ask whether every interaction improves a proprietary dataset and whether the workflow is deep enough to resist recreation by the underlying system of record.
Fintech is a favored exception because “if you’re moving money, you’re generally in a good place.” Rajaram sees anything touching money as having a strong moat and being much more defensible than most pure software.
Rajaram’s uncomfortable conclusion was that “pure software companies are hard.” Investors must believe founders can ship at exceptional velocity while demonstrating that their data, models, and operational embedding compound rather than merely keeping pace with foundation-model releases.
6. Vertical AI must replace the stack, not decorate one function
Harry’s provocation was that agents for dentists, chiropractors, auto manufacturers, or support teams can look like “OpenAI, ElevenLabs for” a vertical. Rajaram called them viable but unlikely to become very large if they remain one function; the ambition must be to own the complete vertical stack.
ServiceTitan is his canonical warning: its S-1 showed 32 products, yet Rajaram described it as still a sub-$10 billion company or something like that. By contrast, Robinhood had 13 product lines above $100 million of revenue and Coinbase had 12, reflecting the broader ceiling available to horizontal platforms.
Vertical software can still fit a $200 million–$400 million venture fund because AI expands the addressable pool from software into services and payroll. The target is not only tooling spend but the much larger budgets for BPO and human labor.
Labor displacement follows a sequence: companies first cut outsourced BPO because AI can deliver comparable or better service 20–30% cheaper; next, they decline to replace departing employees; layoffs come later. Harry’s scale check was striking: Goldman Sachs and Barclays each employ more than 30,000 people in India.
7. Legacy software needs a new business, not a cosmetic AI rescue
For highly valued private companies growing around 15% at roughly $300 million ARR, Rajaram sees two paths. Some become zombies, bolt on AI unsuccessfully, and seek PE buyers or mergers at reset prices; the better-led companies “burn the bridges” and create an AI-native product from scratch.
Intercom and Podium were his positive examples, each building new products beyond $100 million within a few years. The mistake is fixating on repairing the legacy business rather than ruthlessly migrating customers—even at lower prices—and abandoning sunk costs.
Harry pressed the valuation math: if Podium’s agent revenue rises from $100 million to $300 million and then $900 million, paying $5 billion today already prices in two years of triples. Rajaram’s answer was that the thesis only works if Podium replaces the entire software stack, captures digital-labor payroll, and participates in payments—not if it remains a billion-dollar point product.
Kingmaking by mega-funds is real because an eye-opening early valuation can signal quality and attract additional capital, but “just because you king-make doesn’t mean they are the king.” Execution still determines the outcome, and a $10 billion fund is playing a fundamentally different game from a $400 million fund.
8. Seats survive for access; completed work becomes the billing unit
Seat pricing will not disappear because enterprise buyers value predictability. ChatGPT Enterprise and Figma can still sell tiered seats, but each seat must bundle more capability because headcount alone no longer guarantees expansion revenue.
The model breaks when software performs work rather than grants access. Rajaram separates “access products,” appropriately sold by seat, from “work products,” which he guessed should charge for outcomes; he speculated that a legal product such as Harvey might monetize contracts processed rather than users who logged in.
The deeper strategic implication is that pricing must follow the actual constraint. If 100 users process no contracts, the delivered value may be zero; when agents perform the labor, work output—not human access—becomes the economically coherent meter.
9. Revenue velocity matters less than retention tested under pressure
Going from $1 million to $10 million remains excellent, but it no longer produces the “jaw-dropping awe” it once did. The new underwriting question is durability, not early margins: are customers retained, and does revenue expand after initial experimentation?
Rajaram’s rough recollection of Jasper captured the failure mode: it rose from $1 million to $40 million and fell toward $10 million—or perhaps from $1 million to $100 million and back to $40 million. The exact figures were uncertain; the point was that prosumer “tire kickers” can make headline growth evaporate.
He would take triple-triple-double-double growth with excellent gross retention and net revenue retention over 10x growth paired with poor customer retention and NRR below 100%, especially below 90%. Retention should also be evaluated against a “seismic event” such as serious bundled competition; if a company has not faced one, the numbers should be taken with a grain of salt.
Granola and Gamma may be stronger because remarkable products opened non-consumption markets: users paid separately for note-taking or presentations despite bundled alternatives from Zoom, Gmail, or Google Slides. Rajaram nevertheless reiterated that a standalone product eventually needs a second product, discussing Gamma’s possible expansion into documents or websites.
10. Pricing power, market creation, and entry valuation require different tests
Falling inference costs should mechanically improve AI margins, but Rajaram would rather see expansion through pricing power. PayPal reportedly raised prices five times in three years because customers were so sticky; the relevant early-stage question is not a speculative year-five margin, but whether the product might eventually support price increases.
Low-ARPU businesses can work, but only with massive scale and a free or nearly free hook such as Robinhood’s commission-free trading. Selling to wealthy customers or giant enterprises can be structurally easier: Rajaram contrasted millions of consumers with Palantir’s fewer-than-1,000 customers, perhaps fewer, and recalled Veeva going public with 400 customers.
Market sizing still requires bottom-up segmentation, customer interviews, and willingness-to-pay analysis. Rajaram’s Shopify miss came from counting existing e-commerce merchants rather than seeing a platform that could make anyone a seller; non-consumption bets can become the largest wins—or fail completely if the new behavior never emerges.
Prior losses must not dictate the next decision. His model was Mike Moritz backing Instacart after losing $370 million on Webvan in the same broad category: a “paradoxical” demonstration of first-principles thinking. Facebook’s real-identity design supplied a similar differentiator despite appearing to be merely the 52nd social network.
11. Venture returns demand stage discipline, concentration, and timely liquidity
At seed and Series A, Rajaram believes price “almost doesn’t matter” when conviction is right: he entered Faire’s seed round at a $20 million valuation and thinks it has returned roughly 100–200x for him. At Series B and beyond, price can destroy returns; one security company grew from $100 million to $500 million of revenue while remaining valued around the same $4 billion entry price.
Harry described the current Series A market as $3 million–$4 million revenue companies priced at $300 million–$500 million with $30 million–$50 million rounds. Rajaram said this cannot support an entire traditional A fund. He advocates double-digit ownership, patience, and a mixture of A rounds with seed or incubation bets; his fund has 35% reserves.
Reserves reflect an explore-versus-exploit choice. Founder Collective earned an extraordinary Trade Desk multiple from its first check, while IA Ventures repeatedly invested and generated greater dollar returns; Rajaram favors doubling down because one or two companies usually drive a fund, though Harry noted how easily a supposed Hopin or Clubhouse fund can later become the Linear fund.
Liquidity should be judged on forward IRR, not MOIC alone: a 7x return over 20 years can produce only a teens IRR. For an asset representing 20%–40% of a fund, Rajaram feels an obligation to sell some when forward returns are no longer compelling, favoring Fred Wilson’s “sell a third, hold a third, and trade a third” framework once fully liquid.
12. The best defense against stale judgment is proximity to builders
“Proprietary founder access” is not a credible differentiator by itself. A venture firm must offer something distinct—distribution, recruiting, customer access, or judgment—and LPs should verify it by asking founders from the last five companies invested in why they selected that manager and what alternatives they had.
Rajaram learns primarily from entrepreneurs and maintains relationships with people inside model companies to understand “where the turrets or the guns are pointed.” His Google analogy: directly in the roadmap, a startup gets flattened by an implacable tank; even 10 degrees aside, the incumbent may struggle to turn.
His Vanta regret came from having already committed to another company in the space before surveying the category. Meeting Christina convinced him she would win, but he had not scanned all the companies; the lesson was to compare three or four companies first, because the strongest founder’s depth becomes visible against peers.
Quince produced the same correction at the category level. He dismissed it near a $100 million valuation because D2C was unfashionable, missing a 35%–40% repeat-purchase rate before it later raised at $10 billion: “You can’t just take an industry and say it’s good or bad.”
13. In-person alignment and AI-native youth shape the next founder cohort
Rajaram used to believe a fully remote early-stage company could scale with the right culture; he no longer does. He watched companies die because distributed founders could not agree, align strategy, or iterate quickly enough, and now considers at least three in-person days necessary, though not five.
His default advice to graduates is still patience: spend two or three years at a strong company, gaining operating experience and a network before founding. “Life is long.” He does not believe dropping out is right for most young people socially and emotionally.
The exception is an unusual cohort of young builders who “live and breathe differently” and are more “AI maxed” than older workers. Rajaram has invested in more dropouts during the last few months than across his previous 15 years of angel investing, while stressing that only some are exceptional.
His closing optimism was that AI has unlocked ambition to tackle humanity’s hardest problems. The historical misses reinforce the upside: Fitbit returned roughly 500–1,000x at IPO, while even inside Google and Facebook he could not imagine companies debated at roughly $20 billion–$40 billion becoming trillion-dollar institutions.